Graph neural networks and non-commuting operators

Graph neural networks (GNNs) provide state-of-the-art results in a wide variety of tasks which typically involve predicting features at the vertices of a graph. They are built from layers of graph convolutions which serve as a powerful inductive bias for describing the flow of information among the vertices. Often, more than one data modality is available. This work considers a setting in which several graphs have the same vertex set and a common vertex-level learning task. This generalizes standard GNN models to GNNs with several graph operators that do not commute. We may call this model graph-tuple neural networks (GtNN). In this work, we develop the mathematical theory to address the stability and transferability of GtNNs using properties of non-commuting non-expansive operators. We develop a limit theory of graphon-tuple neural networks and use it to prove a universal transferability theorem that guarantees that all graph-tuple neural networks are transferable on convergent graph-tuple sequences. In particular, there is no non-transferable energy under the convergence we consider here. Our theoretical results extend well-known transferability theorems for GNNs to the case of several simultaneous graphs (GtNNs) and provide a strict improvement on what is currently known even in the GNN case. We illustrate our theoretical results with simple experiments on synthetic and real-world data. To this end, we derive a training procedure that provably enforces the stability of the resulting model.

Paper

Similar papers

Peer review

Reviewer RRVw7/10 · confidence 4/52024-07-06

Summary

This paper considers the tasks of extending GNNs to multiple graphs with a shared node set (and the edges representing different types of relations). In order to do this in a principled manner, the authors consider the algebra of non-commutative polynomials and provide a detailed analysis of stability to perturbations. The authors also extend their setting to graphons and show several consistency results (where the practical implications are to transferability). Notably, the proofs are well written with careful attention to detail and do not suffer from many of the issues common to proofs in ML papers. (See e.g., Remark 16.) Overall, this paper provides a solid, easily understandable, theoretical framework for an important problem (learning on multigraphs). Acceptance is recommended.

Strengths

Section 6, connecting the stability guarantees to the training procedure is interesting! Theory is novel, interesting and well explained. The paper is generally well written and well motivated The introduction go WtNNs is interesting, well motivated, and provides a solid theoretical framework for transferability of NNs on multigraphs

Weaknesses

It would be helpful if the sections on non-commutative polynomials and operator filters discussed related work on algebraic signal processing and defining neural nets on arbitrary measure spaces, e.g., https://arxiv.org/pdf/2210.17425, https://arxiv.org/pdf/2108.09923, https://arxiv.org/pdf/2210.16272, https://arxiv.org/pdf/2009.01433, https://arxiv.org/abs/2208.08561 Corollaries 2 and 3 are nice, but should probably be combined into a single corollary. Additionally, it looks like equation 5 could be fit onto a single line (possibly with a small amount of negative hspace.) A more detailed explanation of Example 6 would be helpful. I had to read it several times in order to understand Having some experiments with real data would be helpful (although I do recognize that the contributions of this paper are primarily theoretical). A short conclusion would be nice in the camera ready copy. Line 494: It is not clear where "block operator" norms were introduced In Lemma 10 (and throughout) the authors should make it more clear that \| \cdot \| refers to the \ell^2 norm since there are many different spaces in this paper. Lemma 14 is out of place, it should be before the proof of Lemma 14 (which means it would then become Lemma 13) More details should be added to the proof of theorem 7 Releasing a colab notebook after the review process is okay, but an anonymous github would have been better to improve upon reproducibility Minor: Line 91: "behavior than" should be "behavior as" Line 97: the phrase "loopless" could be confusing. Should make it more clear you mean no self loops The term End(W) is used without definition. I know what this means, but it would be better to discuss in order to keep the paper self contained for an ML audience. Line 167: Cartesian should have a capital C. The authors should check throughout for other words derived from last names, e.g., Euclidean or Laplacian Line 237: V^{(s)} should be V^{(n)}, right? Line 313: "Corollary" should be "corollary" Linne 494: "Lemma" should be "lemma"

Questions

Does any of the graphon theory extend to random graphs derived from graphons? Do you think that analogs of the transferability results for graphon operators can be produced for settings where the graph is derived by subsampling a manifold as in https://arxiv.org/pdf/2305.18467, https://arxiv.org/pdf/2210.00376, https://arxiv.org/pdf/2212.12606, https://arxiv.org/pdf/2307.04056 How does theorem 1 compare with theorem 2 of this paper https://arxiv.org/pdf/2108.09923 Is it possible to define things in any sort of Fourier domain?

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

Yes

Authorsrebuttal2024-08-06

Comments accompanying the rebuttal

**New paragraph referring to related work in the intro** The goal of this paper is to extend the mathematical theory of GNNs to account for multimodal graph settings. The most closely related existing work is the algebraic neural network theory of Parada-Mayorga, Butler and Ribeiro [PMR1][PMR2][PMR3] who pioneer the use of algebras of non-commuting operators. The setting in this paper could be thought of as a special case of this theory. However, there is a crucial difference: whereas the main results in the articles above refer to the Hilbert-Schmidt norm, we introduce and analyze block-operator-norms on non-commutative algebras acting on function spaces. This choice allows us to prove stronger stability and transferability bounds that when restricted to classical GNNs improve upon or complement the state-of-the-art theory. In particular, we complement work in \cite{ruiz2021graph} by delivering bounds that do not exhibit no-transferable energy, and we complement results in \cite{maskey2023transferability} by providing stability bounds that do not require convergence. Our bounds are furthermore easily computable in terms of the networks' parameters improving on the results of~\cite{PMR2} and in particular allow us to devise novel training algorithms with stability guarantees. **New writing of example 6:** Fix $p\in (0,1)$ and let $W(x,y)=p$ for $x\neq y$ and zero otherwise. The graphon Erdos-Renyi graphs $G^{(n)}$ constructed from $W$ as in $(2)$ above are precisely the usual Erdos-Renyi graphs. Part (2) of the sampling Lemma guarantees that $T_W$ converges to $T_{\hat{G}^{(n)}}$ almost surely in the operator norm. By contrast we will show that the sequence **does not converge** to $T_W$ in the Hilbert-Schmidt norm by proving that $\|T_{\hat{G}^{(n)}}(x,y)-T_W(x,y)\|_{\rm HS}>\min(p,1-p)>0$ almost surely. To this end note that for every $n\in \mathbb{N}$ and every $(x,y)\in [0,1]^2$ with $x\neq y$ the difference $|W_{\hat{G}^{(n)}}(x,y)-W(x,y)|\geq \min(p,1-p)$ since the term on the left is either $0$ or $1$. We conclude that $\|T_{\hat{G}^{(n)}}(x,y)-T_W(x,y)\|_{\rm HS} = \|W_{\hat{G}^{(n)}}(x,y)-W(x,y)\|_{L^2([0,1]^2)}\geq \min(p,1-p)>0$ for every $n$ and therefore the sequence fails to converge to zero almost surely. **Short Conclusion (to be added at the end of the article)** In this paper, we introduce graph-tuple networks, a way of extending GNNs to the multimodal graph setting through the use tuples of non-commutative operators endowed with appropriate block-operator norms. We show that GtNNs have several desireable properties such as stability to perturbations and a Universal transferability property on convergent graph-tuples, where the transferability error goes to zero as the graph size goes to infinity. Our transferability theorem improves upon the current state even for the GNN case. Furthermore our error bounds are expressed in terms of computable quantities from the model. This motivates our novel algorithm to enforce stability during training. Experimental results show that our transferability error bounds are reasonably tight, and that our algorithm increases the stability with respect to graph perturbation. In addition, they demonstrate that the transferability theorem holds even for sparse graph tuples. Finally, the experiments on the movie recommendation system suggest that allowing for architectures based on GtNNs is of potential advantage in real-world applications.

Reviewer RRVw2024-08-09

Good Job

I thank the authors for their thorough response.

Reviewer hLUB5/10 · confidence 3/52024-07-12

Summary

This paper introduces a new type of neural network called graph-tuple neural networks (GtNN) that can handle multiple graphs with the same set of nodes and non-commuting graph operators. The authors develop a mathematical theory to show the stability and transferability of GtNNs and derive related bounds, proving a universal transferability theorem for these networks. Their results extend existing transferability theorems for traditional graph neural networks and include experiments on synthetic data to demonstrate their findings.

Strengths

1. The author put forward a novel theoretical framework GtNN to model a general case when considering transferability and stability, and the Universal transferability Theorem to quantify. 2. A generative method of mentioned graph sequence is provided instead of merely assuming the existence of the sequence they used in their theoretical framework. 3. Define the graphon-tuple neural networks (WtNNs) to help theoretical derivation.

Weaknesses

1. No experiments on real data to prove the transferability and stability gain using their method compared to GNN. 2. The theoretical background knowledge is not adequately provided. 3. Lacking proper reasoning phases to connect each sections.

Questions

1. Can this method be viewed as a training strategy that improves every existing Graph Mining method? 2. Since WtNNs are “the natural limits of (GtNNs) as the number of vertices grows to infinity”, why did you discuss WtNNs right after the Perturbation inequalities instead of elaborating on GtNNs first. 3. What is the convergence order, i.e. O(N^k), of the quantity when N goes to infinity?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The authors did not adequate discuss the limitation of their work. They may consider talk about how to implement their complicate theories in a understandable way to help this work to gain influence.

Reviewer eJ8g6/10 · confidence 2/52024-07-14

Summary

This paper considers the node-level learning task with several graphs sharing the same vertex set called graph-tuple neural networks (GtNNs). The stability and transferability of GtNNs are studied using properties of non-commuting non-expansive operators. The authors show that all GtNNs are transferable on convergent graph-tuple sequences using graphon-tuple neural network limits. They also illustrate the theoretical results with experiments on synthetic data.

Strengths

1. A theoretical framework based on operator networks for analyzing graph neural networks on multimodal data where several graphs share the same set of nodes is proposed. Using this framework, a perturbation inequality, which provides the perturbation stability of non-expansive operator-tuple networks, is proven. The proposed framework may be extended to accommodate other settings and can be used to study other tasks related to GNNs. 2. The authors also define the graphon-tuple neural networks (WtNNs) which are natural limits of GtNNs. Further, a Universal transferability Theorem for graphon-graph transference, which guarantees that the GtNN learned on a graph-tuple with sufficient many vertices transfers to other graph-tuples, is shown. This transferability result is original and addresses the multimodal learning aspect in GNNs. 3. This work contributes to the literature of analyzing GNNs through graph limits and the study of the multimodal setting through operator networks is original. The paper is clearly written.

Weaknesses

1. The tightness of the bounds is only demonstrated through numerical experiments rather than formal theoretical analysis. 2. The results are based on strong assumptions such as the existence of graphons induced by the graphs, the continuity of the graphons, the equispaceness of the vertex set, and the graph-tuple convergence in the operator norm. However, these assumptions may be standard in the literature. 3. The results lack support from real-world experiments.

Questions

1. Can the operator network framework be adapted to the case when the graphs have only partial overlap rather than the same set of vertices? 2. Is it possible to extend the current work to *sparse* graphs when there are no induced graphons?

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors adequately addressed the limitations of the work and clearly stated the assumptions.

Reviewer eJ8g2024-08-12

Reply to rebuttal

I thank the authors for the detailed responses and agreeing to address some of the criticism. I maintain my positive score of the paper.

Authorsrebuttal2024-08-06

Anonymous Github repositories

We have added two anonymous Github repositories containing the scripts used for the article: https://anonymous.4open.science/r/GtNN_weighted_circulant_graphs-27EB https://anonymous.4open.science/r/operatorNetworks-508F

Reviewer hLUB2024-08-12

Thank you to the authors for your rebuttal. After reading the rebuttal, I decide to maintain my score.

Authorsrebuttal2024-08-13

Question to reviewers

We thank the reviewers for their time and the constructive and positive feedback we received. Since all of them requested experiments on real data, we'd like to ask them whether the experiment on the movies dataset we provide in this global rebuttal satisfies their request.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC